Individualized credit product intelligent recommendation system for small and micro enterprises

Abstract

The invention discloses a personalized credit product intelligent recommendation system for small and micro enterprises. The system comprises a data acquisition and preprocessing module, a risk assessment and demand analysis module, a credit product knowledge base module, a personalized recommendation algorithm module, a user interaction and feedback module and a system management and monitoring module. The data acquisition module acquires enterprise full-dimension information through multiple interface technologies such as RESTFulAPI, JDBC / ODBC and HTTPS protocols, abnormal value detection is carried out by adopting an isolated forest algorithm, and data preprocessing is carried out by using methods such as a K-nearest neighbor interpolation method and one-hot coding. The risk assessment module adopts a feature screening method combining recursive feature elimination and random forest feature importance to construct a random forest and support vector machine fusion model, and realizes credit risk quantitative scoring through logarithmic function mapping. According to the invention, credit demands of small and micro enterprises can be accurately matched, recommendation accuracy and credit approval efficiency are significantly improved, and credit risks of financial institutions are effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of financial credit technology, and in particular to an intelligent recommendation system for personalized credit products for micro and small enterprises. Background Technology

[0002] Small and micro enterprises (SMEs) occupy an important position in national economic development, playing a vital role in driving economic growth, promoting employment, and stimulating market vitality. However, due to their generally non-standardized operating data, opaque financial information, and weak risk resistance, SMEs face numerous difficulties when applying for bank loans. Traditional financial institutions, often relying on standardized risk assessment systems, struggle to comprehensively and accurately evaluate the true operating conditions and repayment capabilities of SMEs, leading to the long-standing problems of difficult and expensive financing for them. This information asymmetry not only hinders the healthy development of SMEs but also affects the efficiency and quality of financial institutions' services to the real economy.

[0003] Traditional credit product recommendation methods often employ uniform standards and fixed models, failing to provide personalized credit product recommendations based on factors such as the actual operating conditions, credit risk characteristics, and funding needs of micro and small enterprises (MSEs). This "one-size-fits-all" recommendation model leads to a significant mismatch between credit products and the actual needs of MSEs, not only reducing credit approval efficiency and prolonging the time it takes for businesses to obtain funding, but also increasing the credit risk and operating costs of financial institutions. MSEs often need to spend considerable time and effort searching for suitable credit products, while financial institutions struggle to accurately identify high-quality clients, resulting in inefficient resource allocation.

[0004] While some financial institutions have begun to explore data analytics for credit product recommendations, existing solutions still have significant shortcomings in several key areas. Regarding the depth of data processing, current systems often only handle structured financial data, lacking the ability to effectively integrate and analyze unstructured transaction records and industry information. In terms of the accuracy of recommendation algorithms, traditional methods rely primarily on simple rule matching or basic statistical analysis, lacking the ability to deeply mine and learn from complex data relationships. Regarding personalization, existing systems struggle to comprehensively consider the multi-dimensional characteristics and dynamic trends of enterprises, failing to provide truly personalized recommendations that meet the increasingly diverse and differentiated credit needs of micro and small enterprises. Summary of the Invention

[0005] The technical problem this invention aims to solve is that existing micro and small enterprise (MSE) loan product recommendation technologies suffer from insufficient data processing depth, low recommendation algorithm accuracy, and low personalization. These technologies are unable to provide accurate and personalized loan product recommendations based on the multi-dimensional characteristics and actual needs of MSEs, resulting in low matching between loan products and enterprise needs, low loan approval efficiency, and high credit risk for financial institutions.

[0006] To address the aforementioned technical problems, this invention provides an intelligent recommendation system for personalized credit products for micro and small enterprises, employing the following technical solution: The system includes a data acquisition and preprocessing module, a risk assessment and demand analysis module, a credit product knowledge base module, a personalized recommendation algorithm module, a user interaction and feedback module, and a system management and monitoring module.

[0007] The data acquisition and preprocessing module collects comprehensive data from micro and small enterprises through various standardized interfaces. Specifically, it uses a RESTful API to connect with the business registration system, employing OAuth 2.0 authentication to periodically retrieve basic enterprise information; it connects with enterprise ERP systems in real time via JDBC and ODBC interfaces, using data mirroring and synchronization technology to obtain financial data; it obtains tax data through authorized access to the tax bureau's electronic system via HTTPS protocol, supporting XML and JSON data formats; it exports transaction data in batches from the banking system via SFTP file transfer, storing it in CSV format; it obtains credit data from the People's Bank of China Credit Reference Center and third-party credit reporting agencies through encrypted API interfaces, using digital certificate authentication to ensure data security; and it obtains industry data from industry association databases and public websites using the Scrapy crawling framework and Selenium dynamic page parsing technology. In the data preprocessing stage, the Isolation Forest algorithm is used for outlier detection. This algorithm recursively divides the dataset by randomly selecting features and split points, and calculates the outlier score based on the path length of the sample in the tree. For missing value handling, K-nearest neighbor interpolation is used for numerical data, and maximum frequency interpolation is used for categorical data. Data standardization uses Z-score standardization and Min-Max standardization for numerical data, and one-hot encoding and label encoding for categorical data. Data deduplication uses hash deduplication and field combination deduplication strategies.

[0008] The risk assessment and demand analysis module constructs a credit risk assessment model based on machine learning. First, it extracts multi-dimensional features from preprocessed data, including financial status features, operational stability features, credit history features, and industry-related features. A combination of recursive feature elimination and random forest feature importance assessment is used for feature selection. Recursive feature elimination uses a random forest as the base learner, iteratively removing the least important features from the current model, recording the model's performance on the validation set after each feature removal, and selecting the subset of features with the best performance. A random forest model and a support vector machine model are constructed. The random forest model has 100-300 trees, a maximum tree depth of 10-20 layers, and the maximum number of features considered when splitting a node is set to the square root of the total number of features. The support vector machine model selects an appropriate kernel function based on the linear separability of the data, and the penalty parameter C and the RBF kernel function parameter γ are optimized through cross-validation. The stacking method from ensemble learning is used to fuse the prediction results of the two models, using the prediction results of the first layer model as input for training the second layer model. The credit risk quantitative scoring adopts the logarithmic function mapping method, with the formula Score=AB×ln(p / (1-p)), which converts the default probability into a scoring range of 0-100 points and introduces adjustment items such as industry risk adjustment, special event adjustment and trend adjustment.

[0009] The credit product knowledge base module adopts a hybrid storage architecture of knowledge graph and relational database. The relational database MySQL is used to store basic product information, establishing the `product_info` main table, `product_features` feature table, `application_conditions` condition table, and `repayment_methods` repayment method table. A composite index is created for `product_id` and `update_time`, and full-text indexes are added for frequently queried fields. The graph database Neo4j constructs product nodes, attribute nodes, condition nodes, industry nodes, and risk level nodes, establishing semantic relationships between nodes through relation types such as `HAS_FEATURE`, `REQUIRES_CONDITION`, `SUITABLE_FOR`, and `HAS_RISK_LEVEL`. The knowledge update mechanism connects to the financial institution's product release API through API monitoring, using a webhook mechanism to receive update notifications in real time. A Diff algorithm is used to compare old and new versions, updating only the changed parts, and a two-phase commit is used to ensure data consistency.

[0010] The personalized recommendation algorithm module employs a recommendation algorithm based on a combination of collaborative filtering and deep learning. The collaborative filtering algorithm vectorizes the characteristics of micro and small enterprises (MSEs) and the attributes of credit products. Using user-based collaborative filtering, it calculates the similarity between MSEs using the cosine similarity formula sim(i,k)=(xi·xk) / (|xi||xk|), finding the N sets of enterprises most similar to the target enterprise. It then statistically analyzes the credit products these enterprises have access to and generates an initial recommendation list based on the frequency of product occurrence. The deep learning algorithm uses a multilayer perceptron as its basic architecture, comprising an input layer, hidden layers, and an output layer. The input layer concatenates the multi-dimensional data of MSEs and the attribute data of credit products. The hidden layers use multiple fully connected layers with the ReLU activation function, and the output of the l-th layer is h^l=ReLU(W^l×h^(l-1)+b^l). The output layer outputs a scalar representing the matching score. Each credit product in the initial recommendation list, combined with the feature combination of the target enterprise, is input into the deep neural network. Based on the matching score, the initial recommendation list is reordered to generate the final personalized credit product recommendation scheme.

[0011] The user interaction and feedback module provides a visual interactive interface for micro and small enterprises. Enterprises can view recommended credit product information and provide feedback on the recommendation results through this interface. The system collects user feedback information as a basis for optimizing the recommendation algorithm and model, and continuously improves the accuracy of recommendations and user satisfaction.

[0012] The system management and monitoring module allows administrators to manage the system, including setting user permissions, data backup and recovery, system parameter configuration, etc. It monitors the system's operating status in real time, monitors each stage of data collection, processing, and recommendation, and promptly identifies and resolves problems that occur during system operation to ensure stable system operation.

[0013] The beneficial effects of this invention are: First, it achieves precise matching. Through multi-dimensional data collection and in-depth analysis, combined with advanced machine learning algorithms and deep neural networks, it can comprehensively and accurately assess the credit risk and funding needs of micro and small enterprises (MSEs), precisely matching their credit demands with financial institutions' credit products. This significantly improves the matching degree between credit products and MSEs, meeting their diverse financing needs. Compared to traditional standardized recommendation methods, this invention can improve matching accuracy by more than 30%.

[0014] Secondly, it significantly improves efficiency. The automated data collection, preprocessing, and intelligent recommendation processes greatly shorten the time for recommending credit products from several days to within minutes, improving credit approval efficiency and helping micro and small enterprises obtain funding support more quickly. At the same time, financial institutions can also quickly identify high-quality customers and improve business processing efficiency.

[0015] Third, it effectively reduces risk. Scientific risk assessment models comprehensively consider multiple dimensions of factors such as a company's financial situation, operational stability, and credit history. By using a combination of various machine learning algorithms, they can more accurately assess the credit risk of micro and small enterprises, enabling financial institutions to more rationally select credit recipients and determine credit limits, thus effectively reducing credit risk. In practical applications, this can reduce the credit default rate by 15%-20%.

[0016] Fourth, it enhances the user experience. Personalized recommendations and a user-friendly interface enable micro and small enterprises (MSEs) to quickly find the most suitable credit products, reducing the time and effort they spend on product selection, improving their credit application experience, and increasing their satisfaction and loyalty to financial institutions.

[0017] Fifth, it possesses excellent scalability and adaptability. Adopting a modular design architecture, the modules interact with each other through standardized interfaces, facilitating system function expansion and upgrades. The hybrid storage architecture of knowledge graphs and relational databases can flexibly meet the storage needs of different types of credit products. The system can promptly update recommendation strategies and evaluation models based on market changes and policy adjustments, maintaining the system's timeliness and accuracy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the data acquisition and preprocessing process of the present invention; Figure 3 This is a schematic diagram of the risk assessment model construction process of the present invention; Figure 4 This is a schematic diagram of the personalized recommendation algorithm of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0022] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0023] This invention provides an intelligent recommendation system for personalized credit products for micro and small enterprises. This system uses big data analysis and artificial intelligence algorithms to accurately identify the credit needs of micro and small enterprises and recommend personalized products. The specific embodiments of this invention are described in detail below.

[0024] For the overall system architecture, please refer to [link / reference]. Figure 1 As shown The intelligent recommendation system of this invention adopts a modular design, mainly including a data acquisition and preprocessing module, a risk assessment and demand analysis module, a credit product knowledge base module, a personalized recommendation algorithm module, a user interaction and feedback module, and a system management and monitoring module. Each module interacts with data through standardized interfaces, forming a complete intelligent recommendation system.

[0025] participate Figure 2As shown, the data acquisition and preprocessing module serves as the system's data entry point, responsible for collecting comprehensive information on micro and small enterprises from multiple channels. This module interfaces with the business registration system via a RESTful API, employing OAuth 2.0 authentication to periodically retrieve and update basic enterprise information. For financial data from the enterprise's ERP system, JDBC and ODBC interfaces are used for real-time integration, ensuring data timeliness through data mirroring and synchronization technology. Tax data is obtained through authorized access to the tax bureau's electronic system, transmitted using HTTPS, and supports XML and JSON data formats. Bank transaction data is exported in batches via SFTP file transfer and stored in CSV format. Credit data is obtained from the People's Bank of China Credit Reference Center and third-party credit reporting agencies through encrypted API interfaces, using digital certificate authentication to ensure data security. Industry data is obtained from industry association databases and public websites using the Scrapy crawling framework and Selenium dynamic page parsing technology.

[0026] In the data preprocessing stage, the system employs the Isolation Forest algorithm for outlier detection. This algorithm recursively partitions the dataset by randomly selecting features and split points, calculating anomaly scores based on the path length of samples within the tree; shorter paths indicate a higher probability of an anomaly. For missing values, numerical data uses K-nearest neighbor interpolation, filling in the missing value by calculating the mean of the K most similar samples. Categorical data uses the maximum frequency method, filling in the missing value with the most frequent value of that feature. For data standardization, numerical data uses the Z-score standardization formula x′=(x-μ) / σ or the Min-Max standardization formula x′=(x-min(x)) / (max(x)-min(x)). For categorical data, textual data such as business scope and industry category uses one-hot encoding, converting each category into a binary vector; non-numerical data with sequential relationships, such as enterprise credit ratings, uses label encoding, assigning a unique integer value to each category. Data deduplication strategies include hash deduplication and field combination deduplication, identifying duplicate records by calculating the hash value of data records or comparing key field combinations.

[0027] See Figure 3As shown, the risk assessment and demand analysis module constructs a credit risk assessment model for micro and small enterprises using machine learning algorithms. The feature extraction stage extracts key features reflecting credit risk from the preprocessed data, including financial characteristics such as the debt-to-asset ratio (total liabilities / total assets), current ratio (current assets / current liabilities), gross profit margin, and net profit margin; operational stability characteristics such as the company's operating years, the revenue growth rate over the past three years (current year's revenue - previous year's revenue) / previous year's revenue, customer concentration, and supplier concentration; credit record characteristics such as the number of overdue payments, overdue days, current debt amount, and credit score in bank credit records, and the number of overdue tax payments and tax arrears in tax data; and industry-related characteristics such as the industry average debt-to-asset ratio, industry average gross profit margin, and industry prosperity index.

[0028] Feature selection employs a combination of recursive feature elimination and random forest feature importance assessment. Recursive feature elimination uses a random forest as the base learner, iteratively removing the least important features from the current model each time, starting from all features, until a preset number of features is reached or a specific condition is met. During iteration, the model's performance on the validation set is recorded after each feature removal, and the subset of features with the best performance is selected. Random forest feature importance assessment measures the importance score of each feature based on the reduction in impurity caused by node splitting. A threshold is set to retain features with importance scores above the threshold. The intersection of this subset with the recursive feature elimination result yields the final feature subset.

[0029] In terms of model construction, the random forest model sets the number of trees to between 100 and 300, and the maximum tree depth to 10-20 layers to prevent overfitting. The maximum number of features considered when splitting nodes is set to "sqrt" (the square root of the total number of features), the minimum number of splits per sample is set to 2-10, and the minimum number of leaf nodes per sample is set to 1-5. The preprocessed and feature-selected data is divided into a training set (approximately 70%) and a test set (approximately 30%). The random forest model is trained using the training set. For the support vector machine model, an appropriate kernel function is selected based on the linear separability of the data. The penalty parameter C typically ranges from 0.1 to 100, and the parameter γ in the RBF kernel function ranges from 0.001 to 10. Both are optimized through cross-validation.

[0030] To improve the accuracy and stability of the model's predictions, a stacking approach from ensemble learning is used to fuse random forest and support vector machine models. The predictions of both models on the test set are used as new features, combined with the original test set data, and then used as input to a second-layer model, such as a logistic regression model, for training. The second-layer model learns how to better combine these predictive information based on the predictions of the first-layer model and the original data, thereby deriving the final quantitative credit risk score for micro and small enterprises.

[0031] The credit risk quantitative scoring adopts a 100-point scoring system, dividing the credit risk of micro and small enterprises into a range of 0-100 points, with higher scores indicating lower credit risk. Based on the default probability predicted by the model, the probability value is converted into a score through a logarithmic function mapping method. The mapping function is Score=AB×ln(p / (1-p)), where Score is the final credit score, p is the default probability predicted by the model, and A and B are constants that can be determined through calibration using historical data. Adjustment items are introduced on the basic score, including industry risk adjustment, special event adjustment, and trend adjustment. Finally, the basic score and all adjustment items are added to obtain the final credit risk quantitative score for micro and small enterprises.

[0032] The credit product knowledge base module adopts a hybrid storage architecture of knowledge graph and relational database. The relational database MySQL is used to store basic product information, employing the EAV schema to store dynamic attributes. It establishes tables such as `product_info` (main table), `product_features` (feature table), `application_conditions` (condition table), and `repayment_methods` (repayment method table), creating composite indexes for `product_id` and `update_time`, and adding full-text indexes for frequently queried fields. The graph database Neo4j constructs product nodes, attribute nodes, condition nodes, industry nodes, and risk level nodes, establishing semantic relationships between nodes through relation types such as `HAS_FEATURE`, `REQUIRES_CONDITION`, `SUITABLE_FOR`, and `HAS_RISK_LEVEL`. The knowledge update mechanism includes manual and automatic updates. It connects to the financial institution's product release API through API monitoring, uses webhooks to receive update notifications in real time, uses a Diff algorithm to compare old and new versions and only updates the changed parts, and employs a two-phase commit to ensure data consistency.

[0033] See Figure 4 As shown, the personalized recommendation algorithm module employs a recommendation algorithm based on a combination of collaborative filtering and deep learning. The collaborative filtering algorithm vectorizes the characteristics of micro and small enterprises and the attributes of credit products. It uses user-based collaborative filtering to calculate the similarity between micro and small enterprises, employing the cosine similarity formula sim(i,k)=(xi·xk) / (|xi||xk|). For a target micro and small enterprise, it finds the set of N most similar micro and small enterprises, counts the credit products obtained by these enterprises, and generates an initial recommendation list based on the frequency of product occurrence.

[0034] The deep learning algorithm uses a multilayer perceptron as its basic architecture, comprising an input layer, hidden layers, and an output layer. The input layer concatenates multi-dimensional data of micro and small enterprises (MSEs) with attribute data of credit products as the input vector. The hidden layer uses multiple fully connected layers and introduces a non-linear transformation using the ReLU activation function; the output of the l-th layer is h^l = ReLU(W^l × h^(l-1) + b^l). The output layer outputs a scalar representing the matching score between the target MSE and the credit product. Each credit product in the initial recommendation list, combined with the feature combination of the target MSE, is input into the deep neural network. Based on the matching score, the initial recommendation list is reordered to generate a personalized credit product recommendation scheme.

[0035] The user interaction and feedback module provides a visual interface for micro and small enterprises (MSEs). Enterprises can view recommended credit product information and provide feedback on the recommendations. The system collects user feedback as a basis for optimizing the recommendation algorithm and model. The system management and monitoring module allows administrators to perform management operations such as setting user permissions, data backup and recovery, and system parameter configuration. It also monitors the system's operational status in real time, tracking data collection, processing, and recommendation processes.

[0036] Example 1 Taking the application for credit product recommendation by a small and micro manufacturing enterprise as an example, the actual application process of the system of the present invention will be explained in detail.

[0037] Established in 2018 with a registered capital of 2 million yuan, this manufacturing company primarily engages in the processing of mechanical parts and employs 30 people. The system first acquires comprehensive information about the company through its data acquisition and preprocessing module. Basic information, including establishment date, registered address, business scope, and registered capital, is obtained from the business registration system. Financial data for the past three years is obtained through integration with the company's ERP system, including a balance sheet showing total assets of 5 million yuan, total liabilities of 2 million yuan, and a debt-to-asset ratio of 40%; an income statement showing operating revenue of 8 million yuan, net profit of 500,000 yuan, and a net profit margin of 6.25%; and a cash flow statement showing net cash flow from operating activities of 800,000 yuan. Tax records from the tax system show that the company pays taxes normally with no overdue records and has a tax credit rating of A. Transaction data from banks shows stable business relationships with 5 major customers and 8 suppliers, with a customer concentration of 60% and a supplier concentration of 45%. Credit inquiries show that the company has no adverse credit records and a credit score of 750.

[0038] During the data preprocessing stage, the system cleans and standardizes the acquired data. An isolated forest algorithm is used to detect and remove individual abnormal transaction data, and K-nearest neighbor interpolation is used to fill in missing financial indicators. The company's business scope, "machinery parts processing," is coded using one-hot encoding, and the industry category is marked as manufacturing. Numerical financial indicators are standardized using Z-scores to ensure that indicators with different dimensions can be analyzed uniformly.

[0039] The risk assessment and demand analysis module conducted a credit risk assessment of the company. The system extracted financial characteristics including a debt-to-equity ratio of 40%, a current ratio of 2.1, and a net profit margin of 6.25%; operational stability characteristics included a 7-year operating history and revenue growth rates of 15%, 12%, and 8% for the past three years, showing a stable growth trend; credit record characteristics showed a credit score of 750 with no overdue records; and industry-related characteristics showed that the average debt-to-equity ratio for the manufacturing industry was 45%, which was lower than the industry average for this company.

[0040] Through recursive feature elimination and random forest feature importance analysis, the system selected the 20 most important features for model training. The enterprise data was input into the trained fusion model. The random forest model predicted a default probability of 5%, the support vector machine model predicted a default probability of 4.5%, and the final default probability was 4.7% obtained through a stacking method. The base score was calculated as 750 points using the mapping function Score = 600 - 50 × ln(0.047 / (1-0.047)). Considering industry risk adjustments, manufacturing is a medium-risk industry, so the adjustment is 0 points; no special events occurred, so the adjustment is 0 points; the positive trend of continuous revenue growth leads to an upward adjustment of 5 points. The final credit risk quantitative score is 755 points, belonging to the lower-level (Level 2) risk category.

[0041] The credit product knowledge base module contains information on 200 credit products from 50 financial institutions. The system filters eligible credit products from the knowledge base based on factors such as the enterprise's risk level, funding needs, and industry attributes. Through graph database queries, 15 credit products suitable for the manufacturing industry with a risk level requirement of level two or below were found, including ICBC's "Small and Micro Enterprise Working Capital Loan," CCB's "Growth Path Credit Loan," and CMB's "Small Business Credit Loan," among others.

[0042] The personalized recommendation algorithm module first uses a collaborative filtering algorithm to find other companies similar to the manufacturing company. Through cosine similarity calculation, 10 similar companies are found, which are highly similar to the target company in terms of asset size, industry attributes, and financial status. The credit products available to these 10 companies are analyzed, revealing that 6 companies chose the Industrial and Commercial Bank of China's "Small and Micro Enterprise Working Capital Loan," 4 companies chose the China Construction Bank's "Growth Path Credit Loan," and 3 companies chose the China Merchants Bank's "Small Business Credit Loan." Based on this, an initial recommendation list is generated.

[0043] The deep learning algorithm concatenates the feature vector of the manufacturing enterprise with the attribute vector of each candidate credit product and inputs it into a trained deep neural network. The network undergoes nonlinear transformations through multiple hidden layers to learn the complex relationship between enterprise features and product attributes. For the Industrial and Commercial Bank of China's (ICBC) "Small and Micro Enterprise Working Capital Loan," the matching score is 0.92; for China Construction Bank's "Growth Path Credit Loan," the matching score is 0.89; and for China Merchants Bank's "Small Business Credit Loan," the matching score is 0.85. Based on the matching scores, the initial recommendation list is reordered, and the final recommendation scheme is as follows: First recommendation: ICBC's "Small and Micro Enterprise Working Capital Loan," with a loan amount of 1-5 million yuan, an annual interest rate of 4.8%-5.5%, and a term of 1-3 years; Second recommendation: China Construction Bank's "Growth Path Credit Loan," with a loan amount of 500,000-3 million yuan, an annual interest rate of 5.0%-5.8%, and a term of 1-2 years; Third recommendation: China Merchants Bank's "Small Business Credit Loan," with a loan amount of 300,000-2 million yuan, an annual interest rate of 5.2%-6.0%, and a term of 1 year.

[0044] The user interaction and feedback module presents the recommended results to the manufacturing company through a visual interface. The interface displays detailed information for each recommended product, including loan amount, interest rate range, repayment method, application requirements, etc., and provides a product comparison function. After reviewing the recommendations, the company's management expressed satisfaction with the top-recommended product and submitted a loan application, while also providing positive feedback on the accuracy of the system's recommendations. The system records this feedback and uses it for subsequent algorithm optimization to improve recommendation accuracy.

[0045] Through the system management and monitoring module, administrators can view the system's operational status in real time, including key indicators such as data collection success rate, model prediction accuracy, and user satisfaction. In this recommendation process, the data collection completion rate was 98%, the risk assessment accuracy rate was 95%, the recommended product matching rate was 90%, and the user satisfaction score was 4.5 out of 5. All indicators met expectations, validating the system's effectiveness and practicality.

[0046] This embodiment fully demonstrates the complete process and significant effects of the intelligent recommendation system of the present invention in practical applications. Through multi-dimensional data analysis, scientific risk assessment, and intelligent recommendation algorithms, it successfully provides accurate and personalized credit product recommendation services for micro and small enterprises, significantly improving the matching degree between credit products and enterprise needs, increasing credit approval efficiency, and reducing the credit risk of financial institutions.

[0047] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0048] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A personalized credit product intelligent recommendation system for micro and small enterprises, characterized in that, include: The data acquisition and preprocessing module uses a RESTful API interface to connect with the business registration system and adopts the OAuth2.0 authentication mechanism to obtain basic enterprise information. It connects with the enterprise ERP system in real time through JDBC and ODBC interfaces to obtain financial data. It obtains tax data by authorizing access to the tax bureau's electronic system through HTTPS protocol. It exports transaction data in batches from the bank system through SFTP file transfer. It obtains credit data from credit reporting agencies through encrypted API interfaces. It uses the isolated forest algorithm for outlier detection, the K-nearest neighbor interpolation method to handle missing values ​​in numerical data, and one-hot coding and label coding methods to standardize categorical data. The risk assessment and demand analysis module uses a combination of recursive feature elimination and random forest feature importance to screen features, constructs a random forest model and a support vector machine model, uses a stacking method to fuse the prediction results of the two models, and converts the default probability into a credit risk quantitative score through a logarithmic function mapping method. The credit product knowledge base module uses a MySQL relational database to store basic product information and a Neo4j graph database to build semantic relationships between product nodes, attribute nodes, condition nodes, industry nodes, and risk level nodes. It receives product update notifications in real time through API monitoring and Webhook mechanisms. The personalized recommendation algorithm module uses a user-based collaborative filtering algorithm to calculate the similarity between micro and small enterprises using cosine similarity to generate an initial recommendation list. Then, a multilayer perceptron deep neural network is used to optimize and sort the initial recommendation list to generate the final recommendation scheme. The user interaction and feedback module provides a visual interactive interface for micro and small enterprises to view recommendation results and provide feedback. The system management and monitoring module manages and configures the system and monitors its operational status in real time.

2. The intelligent recommendation system according to claim 1, characterized in that, The isolated forest algorithm recursively divides the dataset into two subsets by randomly selecting features and split points, and calculates anomaly scores based on the path length of the samples in the tree.

3. The intelligent recommendation system according to claim 1, characterized in that, The recursive feature elimination uses random forest as the base learner. Starting from all features, each iteration deletes the least important features in the current model, records the model's performance on the validation set after each feature deletion, and selects the feature subset with the best performance.

4. The intelligent recommendation system according to claim 1, characterized in that, The random forest model sets the number of trees to between 100 and 300, the maximum depth of the trees to between 10 and 20 layers, the maximum number of features considered when splitting a node to the square root of the total number of features, the minimum number of splits for a sample to be 2-10, and the minimum number of leaf nodes for a sample to be 1-5.

5. The intelligent recommendation system according to claim 1, characterized in that, The support vector machine model selects a linear kernel function, radial basis function, or polynomial kernel function based on the linear separability of the data. The penalty parameter C ranges from 0.1 to 100, and the parameter γ in the RBF kernel function ranges from 0.001 to 10.

6. The intelligent recommendation system according to claim 1, characterized in that, The logarithmic function mapping method uses the formula Score=AB×ln(p / (1-p)), where Score is the final credit score, p is the default probability predicted by the model, and A and B are constants determined through calibration using historical data.

7. The intelligent recommendation system according to claim 1, characterized in that, The MySQL relational database includes the product_info main table, the product_features feature table, the application_conditions condition table, and the repayment_methods repayment method table, with a composite index created for product_id and update_time.

8. The intelligent recommendation system according to claim 1, characterized in that, The cosine similarity calculation formula is sim(i,k)=(xi·xk) / (|xi||xk|), where xi and xk are the feature vectors of the i-th and k-th micro-enterprises, respectively.

9. The intelligent recommendation system according to claim 1, characterized in that, The multilayer perceptron deep neural network includes an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function, and the output of the l-th layer is h^l=ReLU(W^l×h^(l-1)+b^l), where W^l is the weight matrix of the l-th layer and b^l is the bias vector.

Images